342 lines
14 KiB
Plaintext
342 lines
14 KiB
Plaintext
//
|
|
// Generated by NVIDIA NVVM Compiler
|
|
//
|
|
// Compiler Build ID: CL-23083092
|
|
// Cuda compilation tools, release 9.1, V9.1.85
|
|
// Based on LLVM 3.4svn
|
|
//
|
|
|
|
.version 6.1
|
|
.target sm_30
|
|
.address_size 64
|
|
|
|
// .globl _Z11interpolatei
|
|
.global .align 8 .b8 pixelID[8];
|
|
.global .align 8 .b8 resolution[8];
|
|
.global .align 4 .b8 normal[12];
|
|
.global .align 4 .b8 camPos[12];
|
|
.global .align 4 .b8 root[4];
|
|
.global .align 4 .u32 imageEnabled;
|
|
.global .texref lightmap;
|
|
.global .align 16 .b8 tileInfo[16];
|
|
.global .align 4 .u32 additive;
|
|
.global .align 8 .b8 texCoords[8];
|
|
.global .align 4 .f32 triAlpha;
|
|
.global .align 1 .b8 vertex_buffer[1];
|
|
.global .align 1 .b8 index_buffer[1];
|
|
.global .align 4 .u32 firstAlphaTriangle;
|
|
.global .align 1 .b8 vertex_buffer_uv[1];
|
|
.global .align 1 .b8 triangleAlphaIDs[1];
|
|
.global .align 1 .b8 alphaTextures[1];
|
|
.global .align 1 .b8 translucentTexIDs[1];
|
|
.global .align 4 .b8 ray[36];
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo7pixelIDE[8] = {82, 97, 121, 0, 8, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo10resolutionE[8] = {82, 97, 121, 0, 8, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo6normalE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo6camPosE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo4rootE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo12imageEnabledE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo8tileInfoE[8] = {82, 97, 121, 0, 16, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo8additiveE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo9texCoordsE[8] = {82, 97, 121, 0, 8, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo8triAlphaE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo18firstAlphaTriangleE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo3rayE[8] = {82, 97, 121, 0, 36, 0, 0, 0};
|
|
.global .align 8 .u64 _ZN21rti_internal_register20reg_bitness_detectorE;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail0E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail1E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail2E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail3E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail4E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail5E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail6E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail7E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail8E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail9E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail0E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail1E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail2E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail3E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail4E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail5E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail6E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail7E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail8E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail9E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_xE;
|
|
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_yE;
|
|
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_zE;
|
|
.global .align 8 .b8 _ZN21rti_internal_typename7pixelIDE[6] = {117, 105, 110, 116, 50, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename10resolutionE[6] = {117, 105, 110, 116, 50, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename6normalE[7] = {102, 108, 111, 97, 116, 51, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename6camPosE[7] = {102, 108, 111, 97, 116, 51, 0};
|
|
.global .align 16 .b8 _ZN21rti_internal_typename4rootE[9] = {114, 116, 79, 98, 106, 101, 99, 116, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typename12imageEnabledE[4] = {105, 110, 116, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename8tileInfoE[6] = {117, 105, 110, 116, 52, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typename8additiveE[4] = {105, 110, 116, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename9texCoordsE[7] = {102, 108, 111, 97, 116, 50, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename8triAlphaE[6] = {102, 108, 111, 97, 116, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typename18firstAlphaTriangleE[4] = {105, 110, 116, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typename3rayE[4] = {82, 97, 121, 0};
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum7pixelIDE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum10resolutionE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum6normalE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum6camPosE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum4rootE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum12imageEnabledE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum8tileInfoE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum8additiveE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum9texCoordsE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum8triAlphaE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum18firstAlphaTriangleE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum3rayE = 4919;
|
|
.global .align 16 .b8 _ZN21rti_internal_semantic7pixelIDE[14] = {114, 116, 76, 97, 117, 110, 99, 104, 73, 110, 100, 101, 120, 0};
|
|
.global .align 16 .b8 _ZN21rti_internal_semantic10resolutionE[12] = {114, 116, 76, 97, 117, 110, 99, 104, 68, 105, 109, 0};
|
|
.global .align 16 .b8 _ZN21rti_internal_semantic6normalE[17] = {97, 116, 116, 114, 105, 98, 117, 116, 101, 32, 110, 111, 114, 109, 97, 108, 0};
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic6camPosE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic4rootE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic12imageEnabledE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic8tileInfoE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic8additiveE[1];
|
|
.global .align 16 .b8 _ZN21rti_internal_semantic9texCoordsE[20] = {97, 116, 116, 114, 105, 98, 117, 116, 101, 32, 116, 101, 120, 67, 111, 111, 114, 100, 115, 0};
|
|
.global .align 16 .b8 _ZN21rti_internal_semantic8triAlphaE[19] = {97, 116, 116, 114, 105, 98, 117, 116, 101, 32, 116, 114, 105, 65, 108, 112, 104, 97, 0};
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic18firstAlphaTriangleE[1];
|
|
.global .align 16 .b8 _ZN21rti_internal_semantic3rayE[13] = {114, 116, 67, 117, 114, 114, 101, 110, 116, 82, 97, 121, 0};
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation7pixelIDE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation10resolutionE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation6normalE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation6camPosE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation4rootE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation12imageEnabledE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation8tileInfoE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation8additiveE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation9texCoordsE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation8triAlphaE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation18firstAlphaTriangleE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation3rayE[1];
|
|
|
|
.visible .entry _Z11interpolatei(
|
|
.param .u32 _Z11interpolatei_param_0
|
|
)
|
|
{
|
|
.reg .pred %p<5>;
|
|
.reg .f32 %f<99>;
|
|
.reg .b32 %r<60>;
|
|
.reg .b64 %rd<101>;
|
|
|
|
|
|
ld.param.u32 %r6, [_Z11interpolatei_param_0];
|
|
// inline asm
|
|
call (%f12, %f13), _rt_get_triangle_barycentrics, ();
|
|
// inline asm
|
|
cvt.s64.s32 %rd4, %r6;
|
|
mov.u64 %rd8, index_buffer;
|
|
cvta.global.u64 %rd3, %rd8;
|
|
mov.u32 %r7, 1;
|
|
mov.u32 %r8, 12;
|
|
mov.u64 %rd7, 0;
|
|
// inline asm
|
|
call (%rd2), _rt_buffer_get_64, (%rd3, %r7, %r8, %rd4, %rd7, %rd7, %rd7);
|
|
// inline asm
|
|
ld.u32 %r1, [%rd2];
|
|
ld.u32 %r2, [%rd2+4];
|
|
ld.u32 %r3, [%rd2+8];
|
|
ld.global.u32 %r4, [firstAlphaTriangle];
|
|
setp.gt.s32 %p1, %r4, %r6;
|
|
mov.f32 %f98, 0f3F800000;
|
|
sub.f32 %f15, %f98, %f12;
|
|
sub.f32 %f3, %f15, %f13;
|
|
@%p1 bra BB0_3;
|
|
|
|
cvt.s64.s32 %rd29, %r3;
|
|
sub.s32 %r17, %r6, %r4;
|
|
cvt.s64.s32 %rd11, %r17;
|
|
mov.u64 %rd33, triangleAlphaIDs;
|
|
cvta.global.u64 %rd10, %rd33;
|
|
mov.u32 %r10, 4;
|
|
// inline asm
|
|
call (%rd9), _rt_buffer_get_64, (%rd10, %r7, %r10, %rd11, %rd7, %rd7, %rd7);
|
|
// inline asm
|
|
ld.u32 %r5, [%rd9];
|
|
cvt.s64.s32 %rd17, %r1;
|
|
mov.u64 %rd34, vertex_buffer_uv;
|
|
cvta.global.u64 %rd16, %rd34;
|
|
mov.u32 %r16, 8;
|
|
// inline asm
|
|
call (%rd15), _rt_buffer_get_64, (%rd16, %r7, %r16, %rd17, %rd7, %rd7, %rd7);
|
|
// inline asm
|
|
ld.v2.f32 {%f16, %f17}, [%rd15];
|
|
cvt.s64.s32 %rd23, %r2;
|
|
// inline asm
|
|
call (%rd21), _rt_buffer_get_64, (%rd16, %r7, %r16, %rd23, %rd7, %rd7, %rd7);
|
|
// inline asm
|
|
ld.v2.f32 {%f20, %f21}, [%rd21];
|
|
// inline asm
|
|
call (%rd27), _rt_buffer_get_64, (%rd16, %r7, %r16, %rd29, %rd7, %rd7, %rd7);
|
|
// inline asm
|
|
mul.f32 %f24, %f12, %f20;
|
|
mul.f32 %f25, %f12, %f21;
|
|
fma.rn.f32 %f26, %f3, %f16, %f24;
|
|
fma.rn.f32 %f27, %f3, %f17, %f25;
|
|
ld.v2.f32 {%f28, %f29}, [%rd27];
|
|
fma.rn.f32 %f4, %f13, %f28, %f26;
|
|
fma.rn.f32 %f5, %f13, %f29, %f27;
|
|
and.b32 %r18, %r5, -16384;
|
|
setp.eq.s32 %p2, %r18, 16384;
|
|
@%p2 bra BB0_7;
|
|
bra.uni BB0_2;
|
|
|
|
BB0_7:
|
|
add.s32 %r39, %r5, -16384;
|
|
cvt.u64.u32 %rd59, %r39;
|
|
mov.u64 %rd63, translucentTexIDs;
|
|
cvta.global.u64 %rd58, %rd63;
|
|
// inline asm
|
|
call (%rd57), _rt_buffer_get_64, (%rd58, %r7, %r10, %rd59, %rd7, %rd7, %rd7);
|
|
// inline asm
|
|
ld.u32 %r37, [%rd57];
|
|
mov.u32 %r38, 2;
|
|
mov.f32 %f49, 0f00000000;
|
|
// inline asm
|
|
call (%f42, %f43, %f44, %f45), _rt_texture_get_f_id, (%r37, %r38, %f4, %f5, %f49, %f49);
|
|
// inline asm
|
|
cvt.sat.f32.f32 %f50, %f42;
|
|
mul.f32 %f51, %f50, 0f437F0000;
|
|
cvt.rzi.s32.f32 %r40, %f51;
|
|
cvt.sat.f32.f32 %f52, %f43;
|
|
mul.f32 %f53, %f52, 0f437F0000;
|
|
cvt.rzi.s32.f32 %r41, %f53;
|
|
cvt.sat.f32.f32 %f54, %f44;
|
|
mul.f32 %f55, %f54, 0f437F0000;
|
|
cvt.rzi.s32.f32 %r42, %f55;
|
|
shl.b32 %r43, %r41, 8;
|
|
or.b32 %r44, %r43, %r40;
|
|
shl.b32 %r45, %r42, 16;
|
|
or.b32 %r46, %r44, %r45;
|
|
not.b32 %r47, %r46;
|
|
cvt.rn.f32.s32 %f56, %r47;
|
|
st.global.f32 [triAlpha], %f56;
|
|
bra.uni BB0_6;
|
|
|
|
BB0_2:
|
|
and.b32 %r29, %r5, 65535;
|
|
abs.f32 %f32, %f4;
|
|
cvt.rmi.f32.f32 %f33, %f32;
|
|
sub.f32 %f34, %f32, %f33;
|
|
abs.f32 %f35, %f5;
|
|
cvt.rmi.f32.f32 %f36, %f35;
|
|
sub.f32 %f37, %f35, %f36;
|
|
cvt.u64.u32 %rd47, %r29;
|
|
mov.u64 %rd56, alphaTextures;
|
|
cvta.global.u64 %rd36, %rd56;
|
|
// inline asm
|
|
call (%rd35), _rt_buffer_get_64, (%rd36, %r7, %r10, %rd47, %rd7, %rd7, %rd7);
|
|
// inline asm
|
|
ld.u32 %r21, [%rd35];
|
|
mov.u32 %r27, 2;
|
|
// inline asm
|
|
call (%rd41, %rd42, %rd43, %rd44), _rt_buffer_get_id_size_64, (%r21, %r27, %r7);
|
|
// inline asm
|
|
cvt.rn.f32.u64 %f38, %rd41;
|
|
mul.f32 %f39, %f34, %f38;
|
|
cvt.rzi.u32.f32 %r30, %f39;
|
|
cvt.rn.f32.u64 %f40, %rd42;
|
|
mul.f32 %f41, %f37, %f40;
|
|
cvt.rzi.u32.f32 %r31, %f41;
|
|
// inline asm
|
|
call (%rd45), _rt_buffer_get_64, (%rd36, %r7, %r10, %rd47, %rd7, %rd7, %rd7);
|
|
// inline asm
|
|
ld.u32 %r26, [%rd45];
|
|
cvt.u64.u32 %rd52, %r30;
|
|
cvt.u64.u32 %rd53, %r31;
|
|
// inline asm
|
|
call (%rd51), _rt_buffer_get_id_64, (%r26, %r27, %r7, %rd52, %rd53, %rd7, %rd7);
|
|
// inline asm
|
|
ld.u8 %r32, [%rd51];
|
|
shr.u32 %r33, %r5, 16;
|
|
and.b32 %r34, %r32, %r33;
|
|
setp.eq.s32 %p3, %r34, 0;
|
|
selp.f32 %f98, 0f00000000, 0f3F800000, %p3;
|
|
|
|
BB0_3:
|
|
st.global.f32 [triAlpha], %f98;
|
|
cvt.s64.s32 %rd84, %r1;
|
|
mov.u64 %rd100, vertex_buffer;
|
|
cvta.global.u64 %rd65, %rd100;
|
|
mov.u32 %r59, 32;
|
|
// inline asm
|
|
call (%rd64), _rt_buffer_get_64, (%rd65, %r7, %r59, %rd84, %rd7, %rd7, %rd7);
|
|
// inline asm
|
|
ld.f32 %f57, [%rd64+20];
|
|
ld.f32 %f58, [%rd64+16];
|
|
ld.f32 %f59, [%rd64+12];
|
|
cvt.s64.s32 %rd90, %r2;
|
|
// inline asm
|
|
call (%rd70), _rt_buffer_get_64, (%rd65, %r7, %r59, %rd90, %rd7, %rd7, %rd7);
|
|
// inline asm
|
|
ld.f32 %f60, [%rd70+20];
|
|
ld.f32 %f61, [%rd70+16];
|
|
ld.f32 %f62, [%rd70+12];
|
|
cvt.s64.s32 %rd96, %r3;
|
|
// inline asm
|
|
call (%rd76), _rt_buffer_get_64, (%rd65, %r7, %r59, %rd96, %rd7, %rd7, %rd7);
|
|
// inline asm
|
|
mul.f32 %f63, %f12, %f62;
|
|
mul.f32 %f64, %f12, %f61;
|
|
mul.f32 %f65, %f12, %f60;
|
|
fma.rn.f32 %f66, %f3, %f59, %f63;
|
|
fma.rn.f32 %f67, %f3, %f58, %f64;
|
|
fma.rn.f32 %f68, %f3, %f57, %f65;
|
|
ld.f32 %f69, [%rd76+20];
|
|
ld.f32 %f70, [%rd76+16];
|
|
ld.f32 %f71, [%rd76+12];
|
|
fma.rn.f32 %f72, %f13, %f71, %f66;
|
|
fma.rn.f32 %f73, %f13, %f70, %f67;
|
|
fma.rn.f32 %f74, %f13, %f69, %f68;
|
|
mul.f32 %f75, %f73, %f73;
|
|
fma.rn.f32 %f76, %f72, %f72, %f75;
|
|
fma.rn.f32 %f77, %f74, %f74, %f76;
|
|
sqrt.rn.f32 %f78, %f77;
|
|
rcp.rn.f32 %f79, %f78;
|
|
mul.f32 %f80, %f79, %f72;
|
|
mul.f32 %f81, %f79, %f73;
|
|
mul.f32 %f82, %f79, %f74;
|
|
st.global.f32 [normal], %f80;
|
|
st.global.f32 [normal+4], %f81;
|
|
st.global.f32 [normal+8], %f82;
|
|
// inline asm
|
|
call (%rd82), _rt_buffer_get_64, (%rd65, %r7, %r59, %rd84, %rd7, %rd7, %rd7);
|
|
// inline asm
|
|
ld.v2.f32 {%f83, %f84}, [%rd82+24];
|
|
// inline asm
|
|
call (%rd88), _rt_buffer_get_64, (%rd65, %r7, %r59, %rd90, %rd7, %rd7, %rd7);
|
|
// inline asm
|
|
ld.v2.f32 {%f85, %f86}, [%rd88+24];
|
|
// inline asm
|
|
call (%rd94), _rt_buffer_get_64, (%rd65, %r7, %r59, %rd96, %rd7, %rd7, %rd7);
|
|
// inline asm
|
|
setp.lt.f32 %p4, %f84, 0f00000000;
|
|
@%p4 bra BB0_5;
|
|
bra.uni BB0_4;
|
|
|
|
BB0_5:
|
|
add.f32 %f97, %f84, 0f3F800000;
|
|
st.global.v2.f32 [texCoords], {%f83, %f97};
|
|
bra.uni BB0_6;
|
|
|
|
BB0_4:
|
|
ld.v2.f32 {%f87, %f88}, [%rd94+24];
|
|
mul.f32 %f91, %f12, %f85;
|
|
mul.f32 %f92, %f12, %f86;
|
|
fma.rn.f32 %f93, %f3, %f83, %f91;
|
|
fma.rn.f32 %f94, %f3, %f84, %f92;
|
|
fma.rn.f32 %f95, %f13, %f87, %f93;
|
|
fma.rn.f32 %f96, %f13, %f88, %f94;
|
|
st.global.v2.f32 [texCoords], {%f95, %f96};
|
|
|
|
BB0_6:
|
|
ret;
|
|
}
|
|
|
|
|